WP 2008-19
September 2008
Working Paper
Department of Applied Economics and Management
Cornell University, Ithaca, New York 14853-7801 USA
Do Investors Learn About Analyst Accuracy?
Charles Chang, Hazem Daouk and Albert Wang
Do Investors Learn about Analyst Accuracy?
by
Charles Chang 1 , Hazem Daouk 2 , and Albert Wang 3
1
Cornell University, School of Hotel Administration, Ithaca, NY 14853, phone: 607-255-9882, email:
charlesc@cornell.edu.
2
Cornell University, Department of Applied Economics and Management, Ithaca, NY 14853, phone:
607-255-6459, email: hd35@cornell.edu.
3
Cornell University, Department of Applied Economics and Management, Ithaca, NY 14853, phone:
607-255-6459, email: aw266@cornell.edu.
Electronic copy of this paper is available at: http://ssrn.com/abstract=931704
Do Investors Learn about Analyst Accuracy?
ABSTRACT
We study the impact of analyst forecasts on prices to determine whether
investors learn about analyst accuracy. Our test market is the crude oil
futures market. Prices rise when analysts forecast a decrease (increase) in
crude supplies. In the 15 minutes following supply realizations, prices rise
(fall) when forecasts have been too high (low). In both the initial price
action relative to forecasts and in the subsequent reaction relative to
realized forecast errors, the price response is stronger for more accurate
analysts. These price reactions imply that investors learn about analyst
accuracy and trade accordingly.
Electronic copy of this paper is available at: http://ssrn.com/abstract=931704
Our study seeks to investigate the impact of financial analysts’ forecasts on asset
prices. In addition, we explore investors’ ability to learn about analyst performance and
adjust their trading accordingly. Specifically, we address whether the market reacts to
analyst information, whether analyst forecast error affects price, and whether the market
learns over time which analysts are most accurate. If the answer is negative, one might
question whether analyst performance is important from the firm’s perspective. Why spend
good money maintaining strong analysts if the market fails to react to them and your
investors do not learn or care about their accuracy? Several studies of equity market
analysts discuss analyst forecast accuracy. However, equity analysts must study a variety
of stocks and a variety of factors that impact price, making it difficult to compare the
informational advantages of one analyst to those of another. Furthermore, equity studies
executed at the daily (or longer) horizon potentially admit confounding information events.
Many stocks are also difficult or impossible to short sell, potentially skewing return
studies. We select as our test setting the crude oil futures market. The crude oil futures
market centers around a single underlying asset that all analysts seek to forecast. In
addition, crude prices are principally determined by supply and demand, making analyst
supply forecasts directly relevant. Our data allows us to investigate price movements
immediately following information release thus drastically reducing the likelihood of
confounding events. Finally, oil futures may be shorted without additional costs, and
because of the economic size of this market and its unprecedented activity in recent years,
financial analysts’ forecasts of oil supply (and hence equilibrium price) may have a
substantial economic impact. In this simpler but economically important setting, we find
that analyst supply forecasts have an impact on the price of crude oil products. Price
1
significantly reflects forecast errors and does so within 15 minutes following the
announcement of true supply. This impact overrides that which arises from the supply
announcement itself. We also find that more accurate analysts generate a larger reaction,
inducing larger price movements in response to forecasts and larger corrections when
forecast errors are realized. In essence, we find that information generated by analysts is
important to price determination and that the market seems to learn which analysts are best.
Much empirical work has been aimed at examining the sensitivity of markets to
analyst forecasts and forecast errors. Post-announcement price drift is well documented in
the earnings literature. Kanungo (2004) finds that stock prices are more reactive to forecast
errors than to earnings themselves. Stickel (1992) further finds that those analysts with the
best ex-ante reputations have the largest market impact. In his paper, he concludes that
market participants know the identities of the best analysts and respond accordingly.
However, he finds that this is only true in upward revisions and not in downward ones,
perhaps owing to transactional difficulties in shorting. Sorescu and Subrahmanyam (2004)
similarly find that more experienced analysts receive the largest price reactions,
particularly in the long-run, whereas Chen, Francis, and Jiang (2005) determine that both
the accuracy of earnings forecasts and length of track record increase market reactivity
though that study focuses on first-time analysts. Mikhail, Walther, and Willis (2001), on
the other hand, find the opposite. Splitting firms up by coverage, they find that firms
covered by more accurate, more experienced analysts show less pronounced
post-announcement drift. In work related to commodities, Athanassakos and Kalimipalli
(2004) study the smaller, less liquid, natural gas storage market and find that market
participants recognize differences in ability among gas analysts and hence place more
2
emphasis on lead analysts than on other analysts or on the consensus. In summary, though
post-announcement drift is well documented, the issue of whether investors learn about the
accuracy of analysts elicits less agreement.
In our more straight-forward test market, we find that prices react significantly to
forecasts in the period before true supply is announced. We further find that, when true
supply (and hence forecast error) is revealed, prices adjust immediately, principally within
the first 15 minutes following the announcement. These effects are more significant than
reactions to the supply announcement itself. We then measure the sensitivity of prices to
each individual analyst and find that the reaction is larger for firms with higher past
accuracy, implying that the market adjusts its reaction to analyst forecasts based on past
performance. In other words, investors learn which forecasts to follow.
The remainder of this paper is organized as follows: Section I describes our
datasets. Section II proposes an empirical methodology and results. Section III addresses
robustness checks, and Section IV concludes.
I. Data and Markets
Data for this study comes from two sources: price and transaction data is obtained
from Olsen Data & Associates, and Department of Energy (DOE) supply reports and
analyst forecasts are manually collected from Bloomberg. Our data is collected on a
weekly basis and our study spans the period from June 2003 to March 2005, resulting in a
total time series of 96 weeks.
I.A Olsen Data
The specific asset used to measure oil prices is the Light Sweet Crude futures
3
contract traded on NYMEX (ticker: CL). 1 This is the largest, most liquid, and most
price-transparent contract traded on physical commodities. Traded at a nominal contract
size of 1,000 barrels, CL provides for either physical or cash settlement and is more liquid
and more easily traded than crude oil itself. Bid-ask spreads are 2.5 cents (or roughly 5
bp). 2 Olsen Data & Associates provides transaction data including time (by second) and
price for a variety of contract maturities of CL. The specificity of the time stamp allows us
to match weekly forecast errors with trading surrounding the DOE supply change
announcement. We consider here only the contract with the closest expiration, that is the
near-term future, because it exhibits nearly twice the trading activity of farther-term
contracts. 3 Furthermore, the near-term contract also embeds the least interest rate risk as it
has the shortest duration.
I.B Supply and Forecast Data
We collect our supply and forecast data from Bloomberg, which records the DOE’s
“Weekly Petroleum Status Report” (WPSR) when it is released at 10:30 AM each
Wednesday. 4 This report is prepared by the Energy Information Administration and
includes a detailed summary of the supply of petroleum products in the United States
measured in millions of barrels, both in commercial inventories and the Strategic
Petroleum Reserves (SPR). The focus of analyst forecasts, and hence of this study, is on the
former. WPSR supplies are calculated by the DOE based on representative supply, product
stock, process input, and production numbers collected from selected petroleum
companies. 5 For the purposes of our study, we assume the WPSR to deliver an accurate
measure of supply and will henceforth refer to WPSR supplies as “actual”. Analysts’
forecasted supply changes are likewise manually collected from Bloomberg and verified
4
for accuracy. Though most forecasts are finalized and released early in the week, to ensure
the freshness and consistency of these forecasts at the time of forecast error calculation, we
record them as of 10 A.M. on Wednesday.
In all, we investigate forecasts of supply changes in crude oil, gasoline, and
distillates (petrochemicals such as heating oil, certain plastics, etc.) from 17 financial
institutions. Crude oil consists of a mixture of substances which must be separated in order
to become useful products such as gasoline and heating oil. As such, crude oil prices may
be closely related to the supplies of its refined products. For example, when gasoline
supplies are low, the market may foresee excess demand for crude derived from strong
demand from down-stream producers of gasoline. This prompts investors to drive up the
price of crude. Hence, we test price effects resulting from crude oil alone as well as those
resulting from the cumulative effects of crude oil, gasoline, and distillates. Virtually all
analysts in our data set provide forecasts for all three supplies.
Figure 1 illustrates the historic supply levels of the three separately and in
aggregate, scaled to 6/1/2003 levels. It also shows the relationship between these supplies
and the price of CL. Note that each of the three supplies is negatively correlated with CL,
as would be expected from simple economic arguments (supply increases, prices fall).
Their sum is also negatively correlated to CL, a fact that is expounded upon in Figure 2.
We also find that crude oil supplies are negatively correlated with the supplies of both
gasoline and distillates, which is sensible since crude oil must be processed (consumed) in
order to produce gasoline and distillates.
For a clearer picture as to what drives returns, we evaluate the following regression
in which forecast error (FE) is the forecasted change in supply minus the actual change in
5
supply:
rt = α + β i FE i ,t + β gov Gtot ,t + ε
(1)
We regress returns on forecast errors measured relative to the mean forecast ( FE ) 6
of supply i (crude, gasoline, distillates, and the sum of the three) and while controlling for
aggregate supply change (Gtot). The β’s in this regression describe price reactivity to
forecast error. We speculate that these coefficients are positive. For example, suppose
analysts forecast +1000 barrels of supply change. The market should note the increase in
supply and reduce prices. Suppose then that the true supply is revealed to have zero change
in supply, yielding FE=1000. Prices, having already reacted to forecasts, should rise back
to pre-forecast levels. Hence, the relationship between returns and FE would be positive.
Because of the liquidity afforded in this market, we posit that the price reaction to FE
should be nearly immediate. As can be seen in the graph presented in Figure 2, oil prices
react significantly to supply forecast errors and do so in a relatively short window of time
following the DOE announcement. There is very little return activity before the event
window and, following this initial jump, subsequent price movements do not seem to
reverse the initial reaction—returns are relatively flat for the rest of the trading day. As
such, we define “immediate return” (r) as the return between 10:25 AM and 10:45 AM,
encompassing the announcement time and a short period after. This short window suffices
to capture the lion’s share of returns that occurs as a result of forecast error while
substantially reducing the likelihood of other confounding events occurring within the
event window. The timing of events, then, is:
6
10:00AM
10:25AM
Forecast recorded
10:30AM
10:45AM
DOE supply reported, forecast
error (FE) recorded
Immediate return (r) window
The test results presented in Figure 2 confirm our initial intuition that prices are
significantly related to forecast error. We find that coefficients are positive and significant
for crude, gasoline, and distillates, implying that returns are higher (lower) when supplies
are lower (higher) than expected. Importantly, these effects seem not to be overlapping,
indicating that gasoline and distillate supplies deliver additional information regarding
crude oil prices over-and-above that which is communicated by crude alone. R-squared is
nearly three times as high when looking at the sum of all three supplies compared to crude
oil alone. Because of this, results in the paper will focus on the aggregate forecast error for
the three supplies FEtot. Also, note that oil returns react to forecast errors and not to the
actual supply change itself. βgov is not significant and the inclusion of actual change in
supply does not provide additional explanatory power to our regression.
Summary statistics on supply changes and analyst forecast errors appear in Figures
3 and 4. We calculate these statistics for both the change in number of barrels and the
percent change in total inventory. Regardless of the measure used, forecast errors appear to
be close to Normally distributed. In addition, there do not seem to be large outliers that
might otherwise skew regression results. However, close inspection of these graphs finds
that, when measured by number of barrels, crude oil supply changes and forecast errors are
the most diffuse with standard deviations of 3262 (vs. 2061 for gasoline) and 3263 (vs.
7
1996 for distillates), respectively. However, when measured by percent change in total
supply, distillate supplies and forecast errors are most diffuse with standard deviations of
1.893% (vs. 1.008% for gasoline) and 1.652% (vs. 1.041%), respectively. In light of this
finding, we execute all tests in this study using supply change measured in barrels as well
as in percent change.
II. Empirical Methodology and Results
Having determined that prices react to forecast error, it is sensible to conclude that
investors care about the forecast error of analysts and might learn, over time, which
analysts are most adept. We postulate that, if analysts are known to be accurate, investors
will condition price on their forecasts and price reactions to their forecast errors should be
large. The opposite should be true of inaccurate analysts. We test price reactions before
true supply is revealed to determine whether prices conform to analyst information, and
then examine prices after the supply announcement to see if prices equilibrate to reflect
actual supply information. If investors do not learn about the accuracy of analysts from
their past performance, oil prices should not react differently to different analysts’
forecasts.
Specifically, we propose three testing methodologies. We first execute a test of
investor reaction to forecasts and investigate returns prior to the announcement of true
supply, seeking to determine if investors react more strongly to the most accurate analysts’
forecasts ex-ante. Then, we present two additional regression specifications, a two-stage
and a one-stage, that test if more accurate analysts’ forecast errors induce larger price
corrections ex-post. If analysts do not initially react to forecasts ex-ante, one would not
8
expect to see price correction ex-post. In the tests presented below, our default
specification is to pool all forecasts, but controls for firm- and time- fixed effects are also
presented. Results are qualitatively identical unless otherwise noted.
II.A Ex-ante Regression
First, we test the price-conditioning process ex-ante. We hypothesize that more
accurate analysts’ forecasts elicit larger pre-announcement price movement. To test this,
we first regress returns on forecasts in a rolling window of 20 weeks. We define the ex-ante
return ( r%t ) to be the return from open to close on Tuesday. For each company, we regress
the returns in times t to t+19 on forecasts during the same period. In order to ensure a
representative measure for β t we require at least 10 matched pairs of r%t and FE in the
20-week window. This generates, for each firm, a time series of betas describing how
aggressively prices react to that firm’s forecasts. Specifically, we evaluate:
r%t = α + β%t ( Ft ) + ε
Put simply, we are testing if the market responds to forecasts, which are released prior to
the true supply announcement on Wednesday morning. As forecasts can be changed
throughout the course of the week, we investigate the return over the day closest to the
announcement, Tuesday, to ensure that we capture price movements responding to the
most recent forecasts. 7 In this regression, we conjecture that coefficients should be
negative. When analysts forecast increasing supply, the market reacts by pushing prices
down. The more accurate are the analysts, the more reliable this relationship.
We also calculate for each firm at each point in time an ex-ante accuracy measure
Acc as follows:
9
Acct =
t −1
∑
t −10
1
FEt
N
N is the number of dates for which a forecast error is calculable in the period from t-10 to
t-1. Put simply, Acc is the inverse of the moving average absolute value of forecast errors
over the past 10 periods. To ensure consistency and reliability, we include only those
points at which 5 or more past FE’s exist.
Table I shows summary statistics for each of these measures organized into
accuracy quintiles. The most accurate firms are those with the highest Acc, defined as
“Q1”. Looking at the columns labeled “Beta (Ex-ante),” we find that, in aggregate, beta is
negative. This implies that as forecasts increase, prices decrease. Moreover, we find that
more accurate firms tend to have more negative coefficients. That is, prices fall farther in
response to positive forecasts for more accurate analysts. This relationship is nearly
monotonic, especially when measured by percent change, and the difference in average
beta for Q1 and Q5 is significantly negative.
With these preliminary results in mind, we evaluate the two-stage regression as
follows:
β%t = α + γ% Acct + ε
The resulting γ% from this regression represents the relationship between ex-ante accuracy
and the coefficient of market response to F. If γ% is negative, the market conditions prices
more on the forecasts of firms with higher accuracy. As can be seen in Table II Panel A, we
find that prices seem to react more significantly to the forecasts of more accurate analysts
in the period before actual supplies are announced. Coefficients are negative and
10
significantly so, even when controlling for firm-fixed effects, implying that investors react
more (less) as firms improve (decrease) their accuracy over time. Method of forecast
measurement, by barrels or by percent change, does not seem to affect results.
II.B Two-Stage Ex-post Regression
We now repeat the previous two-stage regression using immediate returns and
forecast errors in order to test ex-post effects. If investors condition prices on the forecasts
of the most accurate analysts, prices should correct most aggressively in response to their
forecast errors. To measure this response, we calculate the following regression:
rt = α + βt FEt + ε
The coefficient β represents the sensitivity of oil prices to the forecast error of a given firm
at time t. A positive β implies that ex-ante over-estimation of supplies induces prices to
increase when actual supplies are found to be lower than forecasts (positive forecast error).
From the columns labeled “Beta (Ex-Post)” found in Table I, we can see that, in aggregate,
beta is in fact positive. We also find that more accurate analysts exhibit more positive
coefficients, indicating that more accurate analysts elicit a larger return response to their
forecast errors. The difference in means for Q1 and Q5 is calculated and found to be
significantly positive. Again, this relationship is nearly monotonic, especially when FE is
measured in percent change.
We then evaluate the second stage of our regression as follows:
βt = α + γ Acct + ε
The γ from this regression represents the relationship between accuracy and the coefficient
of market response to FE. If γ is positive, the market reacts more to firms with higher
ex-ante accuracy, as is consistent with our main proposition. Results presented in Table II
11
Panel B confirm this conjecture, as γ is positive and significant. That is, investors respond
with more sensitivity to forecast errors from firms whose analysis has been more accurate
in the past. Findings are the same whether we aggregate based on change in number of
barrels or on percentage change and are robust to controls for both time- and firm- fixed
effects.
II.C One-Stage Ex-post Regression
Finally, we investigate the relationship between analyst forecasts and prices in a
signaling framework. Consider that forecasts (and resulting forecast error) are information
signals that analysts send to investors. We conjecture that these signals, however, are
attenuated by previous accuracy. If a firm’s past analysis has been very inaccurate, their
perceived signal is zero and constant. Their forecasts should not affect returns. In other
words, if past performance is poor, investors ignore subsequent information signals. To
capture this, we regress returns on weighted forecast errors, using our accuracy measure to
attenuate FE. Specifically, we evaluate the following:
rt = α + γ Acctot ,t *( FEt ) + β ( FEt ) + ε
Put simply, we are attempting to measure the marginal impact of accuracy on how the
forecast error signal FE is transmitted to returns. In this setup, if γ is positive and
significant, investors react to more accurate firms more strongly than inaccurate ones. If γ
is positive and β is not, investors condition prices on accuracy-adjusted forecasts and not
on forecasts alone.
As can be seen in Table II Panel C, we find that γ is positive and significant while β
is not. In other words, the accuracy adjusted forecast error is a significant determinant of
price reaction while forecast error itself is not. Investors seem to adjust their perception of
12
forecast error based on the past accuracy of the analysts. As is consistent with the findings
of the rest of this section, investors seem to condition price most closely on the forecasts of
accurate analysts, and as a result, when errors are realized, price correction is also more
pronounced for more accurate analysts. Again, this is true even when controlling for
firm-fixed effects.
III. Robustness Checks
We execute several robustness checks. Unless otherwise noted, none yield
appreciable differences and our conclusions are unaffected. Numerical results and testing
specifics are available upon request.
For all of our regressions, in calculating aggregate FE, we equally weight the three
supplies. As a robustness check, we also try weighting the forecast errors for each supply
by the in-sample betas calculated in Section 1.2, that is, the empirically derived sensitivity
of price to forecast error. Re-running all of our tests, we find that results are not materially
affected. In addition, we re-run all first-stage regressions with the additional inclusion of
the mean forecast error for all firms on the right-hand side. However, we find that the
resulting measures of price reactivity to firm-specific forecasts and forecast errors are still
significantly related to previous accuracy. In fact, t-stats are the same sign throughout, with
slightly lower magnitudes. Qualitative conclusions remain unchanged. We also apply
alternate specifications for the accuracy measure, measuring it as the inverse standard
deviation, inverse variance, and rank order of forecast errors. In each case, we arrive at the
same qualitative finding: more accurate firms induce a larger surprise response.
We furthermore control for autocorrelation. Durbin-Watson coefficients are about
13
1.86, indicating that auto-correlation is not a significant concern in our tests. Nonetheless,
we calculate Newey-West t-stats for our pooled regressions to control for both
auto-correlation and heteroskedasticity and likewise note no important differences in
significance levels or qualitative understanding.
IV. Conclusion
In this paper, we investigate the impact of analyst forecasts and investors’ ability to
learn in a relatively transparent and straight-forward information setting. By focusing our
attention on the crude oil futures market, we study an asset whose value can be directly
connected to analyst information through simple economic arguments. Papers studying
equity market analysts, particularly those that focus on cross-sectional effects, may be
hindered by idiosyncratic components. Study of the oil market is not affected by such
concerns. The market is also liquid and can be traded equally easily in both the long and
short directions. The inability to short as easily as long in the equity market is a potential
factor that generates asymmetric affects. Our test, again, is free from such concerns.
Furthermore, the high-frequency nature of our data allows us to investigate immediate
price reactions and to isolate the market’s response to analysts from other possible
trade-inducing events. As such, we believe that our study avoids some of the confounding
factors that cloud other studies of this kind and lends a unique voice to the discussion.
We find that investors react to forecasts of the aggregate supply of crude oil,
gasoline, and distillates. We also find that prices correct immediately for forecast errors.
We further discover that investors seem to learn which analysts are most accurate and react
most strongly to those analysts, conditioning prices on their forecasts ex-ante and reacting
14
immediately to their forecasting errors ex-post. This holds true in a number of different
testing specifications and through a variety of statistical checks. Left unanswered,
however, is the exact mechanism through which these price reactions are proliferated. One
possible explanation is that those firms that tend to be most accurate also tend to have the
largest trading desks. In that case, larger price impact on the part of accurate firms could be
self-fulfilling, though this would still be consistent with our intuition regarding learning. It
is also curious that crude oil supplies alone do not seem to explain oil futures returns.
Perhaps a more complex set of interactions is necessary to fully explain the relationships
between supply and price examined here. Since gasoline seems to be an important factor,
electricity prices/supplies or other downstream products may have similar explanatory
power. Without addressing these issues, it suffices here to state that we find oil futures
prices to be intimately related to the aggregate supply of oil products and that investors
recognize this relationship. They seem to trade on the forecasts of analysts in this field and
further seem to learn which analysts are most accurate.
15
References
Amoruso, J. Anthony, Jennifer J. Gaver, and Kenneth M. Gaver, 1998, "The Association
Between Stock Splits and Post-Earnings Announcement Drift", Working Paper,
Available at SSRN: http://ssrn.com/abstract=144418
Athanassakos, George and Madhu Kalimipalli, 2003, “Analyst Forecast Dispersion and
Future Stock Return Volatility”, Quarterly Journal of Business and Economics,
Vol. 42.
Athanassakos, George and Madhu Kalimipalli, 2004, “Analyst forecast performance and
price discovery in a futures market: The case of natural gas storage”, Working
Paper, Richard Ivy School of Business, The University of Western Ontario and
Laurier School of Business and Economics, Wilfred Laurier University, Waterloo,
Canada.
Bartov, Eli, 1992, “Patterns in Unexpected Earnings as an Explanation for
Post-Announcement Drift”, The Accounting Review, Vol. 67, No. 3. Page numbers
missing (actually, they’re missing from most of these, and, when they’re there,
sometime you use “pgs.” while other times writing out “pages.”
Bernard, Victor L. and Jacob K. Thomas, 1989, “Post-Earnings-Announcement Drift:
Delayed Price Response or Risk Premium?” Journal of Accounting Research, Vol.
27.
Bildersee, S. John, 1973, “Some Aspects of the Performance of Non-Convertible Preferred
Stocks”, Journal of Finance, Vol.28, Issue 5, pgs.1187-1201.
Bokheyon, B. Baik and Perk Cheolbeom, 2003, “Dispersion of analysts' expectations and
16
the cross-section of stock returns”, Applied Financial Economics, vol. 13, issue 11.
Bonner, Sarah E, Artur Hugon and Beverly R. Walter, 2005, “Investor Reaction to
Celebrity Analysts: The Case of Earnings Forecast Revisions”, AAA 2006
Financial Accounting and Reporting Section (FARS) Meeting Paper Available at
SSRN: http://ssrn.com/abstract=758510
Brown, Lawrence D, 2001, “How important is past analyst forecast accuracy?” Financial
Analysts Journal, Vol.57, Iss. 6; pg. 44.
Brown, D. Lawrence, Artur Hugon and Liyo Lu, January 2006, “Are Two Heads Better
Than One? The Case of Analyst Forecast Accuracy”, Working Paper, Available at
SSRN: http://ssrn.com/abstract=755824
Chen, Qi, Jennifer Francis, and Wei Jiang, 2005, “Investor learning about analyst
predictive ability”, Journal of Accounting and Economics, Vol. 39, No. 1
Chen, Qi and Jiang Wei, 2006, “Analysts’ Weighting of Private and Public Information”,
The Review of Financial Studies, Vol.19, pg 319.
Chen, Xia, 2004, “Analysts’ affiliation, ranking, and the market reaction to stock
recommendations for IPOs”, Working Paper, Saunder School of Business,
University of British Columbia.
Chinn, Menzie David, Michael LeBlanc, and Coibion Olivier, 2005, "The Predictive
Content of Energy Futures: An Update on Petroleum, Natural Gas, Heating Oil and
Gasoline", NBER
Working
Paper
http://ssrn.com/abstract=645275
17
No.
W11033, Available at
SSRN:
Choi, Wonseok and Jung-Wook Kim, "Under reaction, Trading Volume and Post earnings
announcement
Working
drift",
Paper,
Available
at
SSRN:
http://ssrn.com/abstract=287817
Clement, B. Michael and Senyo Y. Tse, 2005, “Financial Analyst Characteristics and
Herding Behavior in Forecasting”, The Journal of Finance 60(1), pg.307-341.
Coimbra, Carlos and Paulo Esteves Soares, 2004, "Oil Price Assumptions in
Macroeconomic Forecasts: Should we Follow Futures Market Expectations?",
OPEC
Review,
Vol.
28,
No.
2,
pp.
87-106,
Available
at
SSRN:
http://ssrn.com/abstract=565122
Cooper, Rick A., Theodore Day, and Craig M. Lewis, 1999, “Following the Leader: A
Study
of
Individual
Analysts
Earnings
Forecasts”
(http://www.nber.org/~confer/99/cfsi99/lewis.pdf)
Desai, Hemang and Prem C. Jain, 1997, “Long-Run Common Stock Returns following
Stock Splits and Reverse Splits”, Journal of Business, Vol. 7, No. 3.
Easterwood, John and Stacey Nut, 1999, “Inefficiency in Analyst’s Earnings Forecasts:
Systematic Misreaction or Systematic Optimism”, The Journal of Finance, Vol.
LIV, No. 5.
Fang, Lily and Ayako Yasudal, 2005, “Analyst Reputation, Underwriting Pressure, and
Forecast Accuracy”, Paper Submitted for Publication, INSEAD – Finance, France
and Wharton Business School, University of Pennsylvania.
Garfinkel, Jon A. and Jonathan S. Sokobin, 2003, "Volume, Opinion Divergence and
18
Returns: A Study of Post-Earnings Announcement Drift", AFA 2004 San Diego
Meetings Working Paper, Available at SSRN: http://ssrn.com/abstract=280913
Hirshleifer, David A, et al, 2006, “Do Individual Investors Drive Post-Earnings
Announcement Drift? Direct Evidence from Personal Trades”, Journal of Finance,
Forthcoming.
Kanungo, P. Rama, 2004, “Security Analysts and Market Reaction: Caveat for
Monitoring”, (http://ideas.repec.org/p/wpa/wuwpfi/0411039.html)
Liang, Bing, 1990, “Price Pressure: Evidence from the “Dartboard” Column”, Journal of
Business, Vol. 72, No. 1.
Liu, Weimin, Norman C. Strong and Xinzhong Xu, 2003, "Post–earnings–announcement
Drift in the UK", European Financial Management, Vol. 9, pp. 89-116, Available
at SSRN: http://ssrn.com/abstract=388488
Marais, Laurentius M., 1989, “Discussion of Post-Earnings-Announcement Drift: Delayed
Price Response or Risk Premium?” Journal of Accounting Research, Vol. 27.
Mikhail, Michael B., Beverly R. Walther and Richard H. Willis, 2001, "The Effect of
Experience on Security Analyst Under reaction and Post-Earnings-Announcement
Drift", Working Paper, Available at SSRN: http://ssrn.com/abstract=276088
O’Brien, Patricia C., 1990, “Forecast Accuracy of Individual Analysts in Nine Industries”,
Journal of Accounting Research, Vol. 28, No.2 pgs. 286-304.
Park, C. and E. Stice, 2000, “Analyst Forecasting Ability and the Stock Price Reaction to
Forecast Revisions”, Review of Accounting Studies, 5: 259-272.
19
Peterson, David R., 1995, “The Information Role of the Value Line Investment Survey:
Evidence from Stock Highlights”, Journal of Financial and Quantitative Analysis,
Vol. 30, Not. 4.
Riehm, Stephen R, 2000, “The Tao of Forecasting Performance Measurement”, The
Journal of Business Forecasting Methods and Systems; vol.19, 3; ABI/INFORM
Global
Sadka, Gil and Ronnie Sadka, 2004, "The Post-Earnings-Announcement Drift and
Liquidity
Risk",
Working
Paper,
Available
at
SSRN:
http://ssrn.com/abstract=487421
Scherbina, D. Anna, 2004, "Analyst Disagreement, Forecast Bias and Stock Returns", HBS
Publishing, Forthcoming, Available at SSRN: http://ssrn.com/abstract=894381
Sorescu,
S
and
A.
Subrahmanyam,
2004,
“The
Cross-Section
of
Analyst
Recommendations”, Working Paper, Mays Business School, Texas A&M and
Anderson Graduate School of Management, UCLA.
Stickel, Scott E., 1992, “Reputation and Performance among Security Analysts”, Journal
of Finance, American Finance Association, vol. 47(5), pages 1811-36.
Tamura, H, 2002, “Individual-Analyst Characteristics and Forecast Error”, Financial
Analysts Journal, pgs.28-35.
1
Intuition and simple arbitrage arguments imply that futures price changes should match
consistently with changes in spot. Chinn, LeBlanc, and Olivier (2005) find that oil futures
are in fact an unbiased estimator of future spot prices and outperform time-series models.
20
Coimbra and Soares (2004) show that futures outperform macroeconomic models as
estimators for future spot prices.
2
The combination of the regular and “e-miNY” contracts (ticker: QM) allow for liquid
trading at both the institutional and individual levels. QM trades a contract size of 500
barrels, allows for cash settlement, and has a bid-ask spread of 4 cents.
3
NYMEX contracts expire at the end of trading 3 business days before the 25th of each
month. Since our tests focuses on trading immediately surrounding supply announcements,
they should not be significantly affected by expiration effects.
4
In any week where there is a holiday, the report is delayed to Thursday at 10.30 AM.
Historical archives of this data are available on the Department of Energy’s website:
www.eia.doe.gov. We get our DOE reports from Bloomberg though they are available to
the public through a number of media outlets.
5
A full description of the supply data collection process may be found at:
http://www.eia.doe.gov/pub/oil_gas/petroleum/data_publications/weekly_petroleum_stat
us_report/historical/2005/2005_12_21/pdf/appendixa.pdf
6
We use median forecast errors as well and arrive at the same findings.
7
Casual observation and anecdotal evidence find that, in most cases, final forecasts are
released on Tuesday. We, however, repeat all tests using the return accumulated over
Monday and Tuesday and find no qualitative differences.
21
Figure 1: Supply Correlations and their Relation to Price
This figure illustrates the relative relationships between the three supplies examined in this
study: crude oil, gasoline, and distillates. Included as well is the price of crude oil, proxied
by the near-term futures contract CL. Crude supply is found to be negatively correlated to
gasoline and distillates. The total of the three supplies as well as crude oil supply alone are
negatively correlated with price.
Scaled Supply Levels vs CL Price Level
2.00
1.80
1.60
CL
Crude
Gasoline
1.40
Distillates
Total
1.20
1.00
Crude
Gasoline
Distillates
Total
CL
Crude
1.00
-0.16
-0.34
0.15
-0.17
Correlation Matrix
Gasoline
Distillates
-0.16
-0.34
1.00
0.28
0.28
1.00
0.60
0.79
-0.01
-0.04
22
2/2005
12/2004
10/2004
8/2004
6/2004
4/2004
2/2004
12/2003
10/2003
8/2003
6/2003
0.80
Total
0.15
0.60
0.79
1.00
-0.12
CL
-0.17
-0.01
-0.04
-0.12
1.00
Figure 2: Price Response to Forecast Error
This figure reports price responsiveness to forecast error (FE) calculated as forecasted supply minus
actual supply. The graph shows the average absolute value of price movements for oil futures (ticker: CL)
on announcement days. Prices react substantially after forecast errors are realized at 10:30 AM, demarked
by the vertical line. Very little price movement is observed before 10:30 and attenuated drift seems
apparrent throughout the rest of the day. The table shows regression results when returns are regressed on
forecast errors (FE) and aggregate supply change (Gtot). They indicate that the immediate return (change
in price from 10:25 to 10:45 AM) responds significantly positively to forecast errors. When supply is
lower (higher) than expected, prices rise (fall). Crude supplies alone exhibit this effect, but aggregate
supply exhibits the strongest effect. Data is collected from June 2003 to March 2005.
Absolute Price Response to Forecast Error
0.80%
0.70%
0.60%
0.50%
0.40%
0.30%
0.20%
0.10%
0.00%
-0.10%
10:20
AM
Coefficient
(P-value)
Coefficient
(P-value)
Coefficient
(P-value)
Coefficient
(P-value)
Coefficient
(P-value)
10:45
AM
11:10
AM
11:35
AM
12:00
PM
12:25
PM
12:50
PM
Price Reactiveness to Supply Changes
(Coefficients reported x 10^6)
FEcrude FEgasolineFEdistillates FEtot
Gcrude
1.622
(0.0004)
2.240
-0.637
(0.0466)
(0.5173)
2.196
1.548
2.948
(0.0000) (0.0163) (0.0001)
2.210
(0.0001)
2.510
(0.0001)
23
1:15 PM 1:40 PM
Gtot
0.303
(0.5571)
DF
87
R^2
0.126
90
0.115
87
0.327
87
0.3464
87
0.3414
Figure 3: Supply Changes
This figure reports summary statistics regarding supply changes over our test period. Though all three supplies
are similarly distributed, when measured in barrels, crude oil appears to be most diffuse. When measured in
percent change of total supply, distillate supply changes appear to be the most diffuse. Supplies are recorded at
10:30 AM each announcement day over the period from June 2003 to March 2005.
Supply Changes (Million Barrels)
25.0%
20.0%
Crude
Gasoline
Distillates
15.0%
10.0%
5.0%
0.0%
-8000
-6000
-4000
-2000
0
2000
4000
6000
8000
Supply Changes (%)
30.0%
25.0%
20.0%
Crude
Gasoline
Distillates
15.0%
10.0%
5.0%
0.0%
-5.00% -3.75% -2.50% -1.25%
Mean
Median
SD
Max
Min
0.00%
1.25%
2.50%
3.75%
DOE Supply Change Data
Crude Oil
Gasoline
%
%
Gross
Gross
297
0.105%
85
0.046%
300
0.102%
500
0.238%
3262
1.150%
2061
1.008%
7900
2.996%
4900
2.263%
-9100
-3.266%
-5700
-2.895%
24
5.00%
Distillates
%
Gross
20
0.037%
200
0.171%
2276
1.893%
6400
5.037%
-6800
-5.191%
Figure 4: Forecast Errors
This figure presents summary statistics regarding forecast errors. Again, when measured in barrels, crude oil
appears to be most diffuse. When measured in percent change of total supply, distillates appear to be the most
diffuse. Distributions are roughly Normally distributed and there are not a substantial number of outliers.
Forecast errors are calculated as the 10:00 AM forecasted supply minus the 10:30 AM actual supply over the
period from June 2003 to March 2005.
Forecast Errors (Million Barrels)
20.00%
15.00%
Crude
Gasoline
Distillates
10.00%
5.00%
0.00%
-9600 -8000 -6400 -4800 -3200 -1600
0
1600 3200 4800 6400 8000 9600
Forecast Errors (%)
20.00%
15.00%
Crude
Gasoline
Distillates
10.00%
5.00%
Mean
Median
SD
Max
Min
Analyst Forecast Errors
Crude Oil
Gasoline
%
%
Gross
Gross
8
-0.002%
89
0.040%
75
0.026%
100
0.048%
3263
1.156%
2123
1.041%
8900
3.130%
6500
3.107%
-11500 -4.361%
-6400
-2.956%
25
4.
80
%
4.
00
%
3.
20
%
2.
40
%
1.
60
%
0.
80
%
0.
00
%
-4
.8
0%
-4
.0
0%
-3
.2
0%
-2
.4
0%
-1
.6
0%
-0
.8
0%
0.00%
Distillates
%
Gross
-13
-0.008%
100
0.084%
1996
1.652%
8300
6.593%
-7400
-6.410%
Table I: Quintile Analysis
This table tests the monotonicity of the relationship between accuracy and price responsiveness. Accuracy
(Acc) is the inverse of the moving average absolute value of forecast errors over the previous 10 periods. Price
responsiveness (Beta) is calculated in a 20-period rolling regression r = a + Beta*F + e. The return r is the
return on the Tuesday preceding the announcement of actual supply levels for ex-ante tests. Betas are found to
be negatively related to accuracy (meaning positive forecasts induce larger reductions in price for more
accurate analysts). For ex-post tests, the return is calculated from 10:25 AM to 10:45 AM on Wednesday
(forecast error is realized at 10:30 AM). Betas are found to be positively related to accuracy (meaning positive
forecast errors induce larger price increases for more accurate analysts). These relationships are found to be
nearly monotonic and the difference between average values for the extreme quintiles is significantly different
from zero. Barrel coefficients reported x10^-6.
Acc
Aggregate
Q1
Q2
Q3
Q4
Q5
Q1-Q5
(P-value)
169.795
228.873
187.140
166.172
145.322
120.982
107.891
(0.000)
Barrels
Beta
(Ex-Ante)
-0.301
-0.565
-0.474
-0.703
0.054
0.193
-0.758
(0.001)
Beta
(Ex-Post)
1.646
1.708
1.743
1.853
1.557
1.360
0.349
(0.000)
26
Acc
30.971
41.808
35.264
30.884
26.028
20.749
21.059
(0.000)
Percent
Beta
(Ex-Ante)
-0.079
-0.169
-0.098
-0.072
0.011
-0.066
-0.103
(0.002)
Beta
(Ex-Post)
0.273
0.348
0.318
0.265
0.217
0.215
0.133
(0.000)
Table II: Regression Analysis
This table presents tests of the effect of previous-period accuracy on price responsiveness. Panel A
looks at forecasts (F). Panels B and C investigate forecast error (FE). Accuracy (Acc) is the inverse of
the moving average absolute value of forecast errors over the previous 10 periods. Panel A presents
results for a two-stage regression. First, we find calculate price responsiveness (Beta) in a 20-period
rolling regression r = a + Beta*F + e where r is the return on the Tuesday preceding the announcement
of actual supply levels. Reported are gammas from the regression Beta = a + Gamma*Acc. Negative
gamma implies that investors condition more on forecasts of more accurate analysts. Panel B is the
same two-stage regression with first stage r` = a + Beta*FE where r` is the return calculated from
10:25 AM to 10:45 AM on Wednesday (forecast error is realized at 10:30 AM). Reported are the
gammas from the regression: Beta' = a + Gamma'*Acc + e. Panel C evaluates the regression r = a+
Gamma*(Acc*FE) + Beta*FE. In Panels B and C, a positive gamma implies that investors condition
more on the forecasts of more accurate analysts. All coefficients reports x10^-3.
Coefficient
(P-value)
DF
R^2
Panel A: Ex-ante Regression Gammas
Barrels
Pooled FirmTimefixed
fixed
-7.850 -8.830 3.979
(0.0001) (0.0001) (0.1501)
788
773
723
0.019
0.208
0.390
Coefficient
(P-value)
DF
R^2
Panel B: Two-stage Ex-post Regression Gammas
Barrels
Percent
Pooled FirmTimePooled FirmTimefixed
fixed
fixed
fixed
3.140
1.449
5.845
7.220
5.294
7.058
(0.0001) (0.0259) (0.0001)
(0.0001) (0.0001) (0.0001)
788
773
723
788
773
723
0.025
0.871
0.905
0.112
0.807
0.874
Coefficient
(P-value)
DF
R^2
Panel C: One-stage Ex-post Regression Coefficients
Barrels
Percent
Pooled
Firm-fixed
Pooled
Firm-fixed
Acc*FE
FE
Acc*FE
FE
Acc*FE
FE
Acc*FE
FE
8.180
0.000
8.852
0.000
9.780 -63.880 10.670 -77.420
(0.0001) (0.8392) (0.0001) (0.9016)
(0.0001) (0.2111) (0.0001) (0.1298)
950
934
934
950
0.259
0.261
0.282
0.240
27
Percent
FirmTimefixed
fixed
-6.800 -5.790 -0.220
(0.0001) (0.0003) (0.9250)
788
773
723
0.023
0.174
0.359
Pooled